1 #define JEMALLOC_TSD_C_
2 #include "jemalloc/internal/jemalloc_preamble.h"
3 #include "jemalloc/internal/jemalloc_internal_includes.h"
5 #include "jemalloc/internal/assert.h"
6 #include "jemalloc/internal/mutex.h"
7 #include "jemalloc/internal/rtree.h"
9 /******************************************************************************/
12 static unsigned ncleanups;
13 static malloc_tsd_cleanup_t cleanups[MALLOC_TSD_CLEANUPS_MAX];
15 #ifdef JEMALLOC_MALLOC_THREAD_CLEANUP
16 __thread tsd_t JEMALLOC_TLS_MODEL tsd_tls = TSD_INITIALIZER;
17 __thread bool JEMALLOC_TLS_MODEL tsd_initialized = false;
18 bool tsd_booted = false;
19 #elif (defined(JEMALLOC_TLS))
20 __thread tsd_t JEMALLOC_TLS_MODEL tsd_tls = TSD_INITIALIZER;
21 pthread_key_t tsd_tsd;
22 bool tsd_booted = false;
23 #elif (defined(_WIN32))
25 tsd_wrapper_t tsd_boot_wrapper = {false, TSD_INITIALIZER};
26 bool tsd_booted = false;
30 * This contains a mutex, but it's pretty convenient to allow the mutex code to
31 * have a dependency on tsd. So we define the struct here, and only refer to it
32 * by pointer in the header.
34 struct tsd_init_head_s {
35 ql_head(tsd_init_block_t) blocks;
39 pthread_key_t tsd_tsd;
40 tsd_init_head_t tsd_init_head = {
41 ql_head_initializer(blocks),
42 MALLOC_MUTEX_INITIALIZER
44 tsd_wrapper_t tsd_boot_wrapper = {
48 bool tsd_booted = false;
52 /******************************************************************************/
55 tsd_slow_update(tsd_t *tsd) {
56 if (tsd_nominal(tsd)) {
57 if (malloc_slow || !tsd_tcache_enabled_get(tsd) ||
58 tsd_reentrancy_level_get(tsd) > 0) {
59 tsd->state = tsd_state_nominal_slow;
61 tsd->state = tsd_state_nominal;
67 tsd_data_init(tsd_t *tsd) {
69 * We initialize the rtree context first (before the tcache), since the
70 * tcache initialization depends on it.
72 rtree_ctx_data_init(tsd_rtree_ctxp_get_unsafe(tsd));
75 * A nondeterministic seed based on the address of tsd reduces
76 * the likelihood of lockstep non-uniform cache index
77 * utilization among identical concurrent processes, but at the
78 * cost of test repeatability. For debug builds, instead use a
81 *tsd_offset_statep_get(tsd) = config_debug ? 0 :
82 (uint64_t)(uintptr_t)tsd;
84 return tsd_tcache_enabled_data_init(tsd);
88 assert_tsd_data_cleanup_done(tsd_t *tsd) {
89 assert(!tsd_nominal(tsd));
90 assert(*tsd_arenap_get_unsafe(tsd) == NULL);
91 assert(*tsd_iarenap_get_unsafe(tsd) == NULL);
92 assert(*tsd_arenas_tdata_bypassp_get_unsafe(tsd) == true);
93 assert(*tsd_arenas_tdatap_get_unsafe(tsd) == NULL);
94 assert(*tsd_tcache_enabledp_get_unsafe(tsd) == false);
95 assert(*tsd_prof_tdatap_get_unsafe(tsd) == NULL);
99 tsd_data_init_nocleanup(tsd_t *tsd) {
100 assert(tsd->state == tsd_state_reincarnated ||
101 tsd->state == tsd_state_minimal_initialized);
103 * During reincarnation, there is no guarantee that the cleanup function
104 * will be called (deallocation may happen after all tsd destructors).
105 * We set up tsd in a way that no cleanup is needed.
107 rtree_ctx_data_init(tsd_rtree_ctxp_get_unsafe(tsd));
108 *tsd_arenas_tdata_bypassp_get(tsd) = true;
109 *tsd_tcache_enabledp_get_unsafe(tsd) = false;
110 *tsd_reentrancy_levelp_get(tsd) = 1;
111 assert_tsd_data_cleanup_done(tsd);
117 tsd_fetch_slow(tsd_t *tsd, bool minimal) {
118 assert(!tsd_fast(tsd));
120 if (tsd->state == tsd_state_nominal_slow) {
121 /* On slow path but no work needed. */
122 assert(malloc_slow || !tsd_tcache_enabled_get(tsd) ||
123 tsd_reentrancy_level_get(tsd) > 0 ||
124 *tsd_arenas_tdata_bypassp_get(tsd));
125 } else if (tsd->state == tsd_state_uninitialized) {
127 tsd->state = tsd_state_nominal;
128 tsd_slow_update(tsd);
129 /* Trigger cleanup handler registration. */
133 tsd->state = tsd_state_minimal_initialized;
135 tsd_data_init_nocleanup(tsd);
137 } else if (tsd->state == tsd_state_minimal_initialized) {
139 /* Switch to fully initialized. */
140 tsd->state = tsd_state_nominal;
141 assert(*tsd_reentrancy_levelp_get(tsd) >= 1);
142 (*tsd_reentrancy_levelp_get(tsd))--;
143 tsd_slow_update(tsd);
146 assert_tsd_data_cleanup_done(tsd);
148 } else if (tsd->state == tsd_state_purgatory) {
149 tsd->state = tsd_state_reincarnated;
151 tsd_data_init_nocleanup(tsd);
153 assert(tsd->state == tsd_state_reincarnated);
160 malloc_tsd_malloc(size_t size) {
161 return a0malloc(CACHELINE_CEILING(size));
165 malloc_tsd_dalloc(void *wrapper) {
169 #if defined(JEMALLOC_MALLOC_THREAD_CLEANUP) || defined(_WIN32)
174 _malloc_thread_cleanup(void) {
175 bool pending[MALLOC_TSD_CLEANUPS_MAX], again;
178 for (i = 0; i < ncleanups; i++) {
184 for (i = 0; i < ncleanups; i++) {
186 pending[i] = cleanups[i]();
197 malloc_tsd_cleanup_register(bool (*f)(void)) {
198 assert(ncleanups < MALLOC_TSD_CLEANUPS_MAX);
199 cleanups[ncleanups] = f;
204 tsd_do_data_cleanup(tsd_t *tsd) {
205 prof_tdata_cleanup(tsd);
208 arenas_tdata_cleanup(tsd);
210 witnesses_cleanup(tsd_witness_tsdp_get_unsafe(tsd));
214 tsd_cleanup(void *arg) {
215 tsd_t *tsd = (tsd_t *)arg;
217 switch (tsd->state) {
218 case tsd_state_uninitialized:
221 case tsd_state_minimal_initialized:
222 /* This implies the thread only did free() in its life time. */
224 case tsd_state_reincarnated:
226 * Reincarnated means another destructor deallocated memory
227 * after the destructor was called. Cleanup isn't required but
228 * is still called for testing and completeness.
230 assert_tsd_data_cleanup_done(tsd);
232 case tsd_state_nominal:
233 case tsd_state_nominal_slow:
234 tsd_do_data_cleanup(tsd);
235 tsd->state = tsd_state_purgatory;
238 case tsd_state_purgatory:
240 * The previous time this destructor was called, we set the
241 * state to tsd_state_purgatory so that other destructors
242 * wouldn't cause re-creation of the tsd. This time, do
243 * nothing, and do not request another callback.
250 test_callback_t test_callback = *tsd_test_callbackp_get_unsafe(tsd);
251 int *data = tsd_test_datap_get_unsafe(tsd);
252 if (test_callback != NULL) {
259 malloc_tsd_boot0(void) {
267 *tsd_arenas_tdata_bypassp_get(tsd) = true;
272 malloc_tsd_boot1(void) {
274 tsd_t *tsd = tsd_fetch();
275 /* malloc_slow has been set properly. Update tsd_slow. */
276 tsd_slow_update(tsd);
277 *tsd_arenas_tdata_bypassp_get(tsd) = false;
282 _tls_callback(HINSTANCE hinstDLL, DWORD fdwReason, LPVOID lpvReserved) {
284 #ifdef JEMALLOC_LAZY_LOCK
285 case DLL_THREAD_ATTACH:
289 case DLL_THREAD_DETACH:
290 _malloc_thread_cleanup();
299 * We need to be able to say "read" here (in the "pragma section"), but have
300 * hooked "read". We won't read for the rest of the file, so we can get away
309 # pragma comment(linker, "/INCLUDE:__tls_used")
310 # pragma comment(linker, "/INCLUDE:_tls_callback")
312 # pragma comment(linker, "/INCLUDE:_tls_used")
313 # pragma comment(linker, "/INCLUDE:tls_callback")
315 # pragma section(".CRT$XLY",long,read)
317 JEMALLOC_SECTION(".CRT$XLY") JEMALLOC_ATTR(used)
318 BOOL (WINAPI *const tls_callback)(HINSTANCE hinstDLL,
319 DWORD fdwReason, LPVOID lpvReserved) = _tls_callback;
322 #if (!defined(JEMALLOC_MALLOC_THREAD_CLEANUP) && !defined(JEMALLOC_TLS) && \
325 tsd_init_check_recursion(tsd_init_head_t *head, tsd_init_block_t *block) {
326 pthread_t self = pthread_self();
327 tsd_init_block_t *iter;
329 /* Check whether this thread has already inserted into the list. */
330 malloc_mutex_lock(TSDN_NULL, &head->lock);
331 ql_foreach(iter, &head->blocks, link) {
332 if (iter->thread == self) {
333 malloc_mutex_unlock(TSDN_NULL, &head->lock);
337 /* Insert block into list. */
338 ql_elm_new(block, link);
339 block->thread = self;
340 ql_tail_insert(&head->blocks, block, link);
341 malloc_mutex_unlock(TSDN_NULL, &head->lock);
346 tsd_init_finish(tsd_init_head_t *head, tsd_init_block_t *block) {
347 malloc_mutex_lock(TSDN_NULL, &head->lock);
348 ql_remove(&head->blocks, block, link);
349 malloc_mutex_unlock(TSDN_NULL, &head->lock);